Skip to content
InfoResearchPreprintLLM-specific

Dressed in Authority: Practical Black-Box Jailbreaking of MLLMs via Visual Trust Bias

Published
Record updated
View JSON

Summary

Researchers introduce Visual Trust Bias Attack (VTBA), a black-box jailbreak for multimodal large language models (MLLMs) that exploits their trust bias toward authoritative visual layouts. VTBA matches each naturally occurring harmful image-text pair to an authoritative template and embeds the harmful intent in it, overriding safety refusals. Across 2,500 image-text pairs, the authors report consistently high attack success rates on strong MLLMs.